Fix 2.8 issue per sample grad #4937

Job	Run time
pytorch_tutorial_build_worker (2, 15, linux.g5.4xlarge.nvidia.gpu)	37m 58s
pytorch_tutorial_build_worker (1, 15, linux.g5.12xlarge.nvidia.gpu)	15m 52s
pytorch_tutorial_build_worker (3, 15, linux.g5.4xlarge.nvidia.gpu)	23m 45s
pytorch_tutorial_build_worker (4, 15, linux.g5.4xlarge.nvidia.gpu)	36m 37s
pytorch_tutorial_build_worker (6, 15, linux.g5.4xlarge.nvidia.gpu)	32m 0s
pytorch_tutorial_build_worker (7, 15, linux.g5.4xlarge.nvidia.gpu)	22m 58s
pytorch_tutorial_build_worker (14, 15, linux.g5.4xlarge.nvidia.gpu)	30m 2s
pytorch_tutorial_build_worker (8, 15, linux.g5.4xlarge.nvidia.gpu)	25m 52s
pytorch_tutorial_build_worker (10, 15, linux.g5.4xlarge.nvidia.gpu)	14m 50s
pytorch_tutorial_build_worker (11, 15, linux.g5.4xlarge.nvidia.gpu)	15m 17s
pytorch_tutorial_build_worker (15, 15, linux.g5.4xlarge.nvidia.gpu)	20m 22s
pytorch_tutorial_build_worker (13, 15, linux.g5.4xlarge.nvidia.gpu)	17m 5s
pytorch_tutorial_build_worker (5, 15, linux.g5.4xlarge.nvidia.gpu)	24m 1s
pytorch_tutorial_build_worker (9, 15, linux.g5.4xlarge.nvidia.gpu)	20m 33s
pytorch_tutorial_build_worker (12, 15, linux.g5.4xlarge.nvidia.gpu)	15m 38s
pytorch_tutorial_build_manager	0s
	5h 52m 50s

Provide feedback